How this play with rumored / leaked Raptor Lake Refresh release ? Intel is silent about that refresh, and other leaks suggest MTL for desktop to be launched anyway. For Intel launch mean 'silicon out of factory' not 'in shops'Reply
Well, the rumors are that MTL will be used primarily for laptop, and *possibly* also for some i5 variants. Not saying that's what's going to happen, but that's what the rumor mill is saying.Reply
So yields are still rather poor and this is just getting pushed out the door to meet executive timelines.
— 1st 14nm CPU was basically only mobile (Broadwell) — 1st 10nm CPU was basically only mobile (Ice Lake and Cannon Lake) — 1st 7nm (Intel 4) CPU will be primarily mobile
See, TSMC does the same ramp (~90 to 110 mm^2 iPhone chips first), but they're humble enough to admit it as normal and part of the process.
Intel? "Oh, God, no. QUICK, a re-brand! Confuse the customers!"Reply
>So yields are still rather poor and this is just getting pushed out the door to meet executive timelines. Again, not to ascribe the rumor mill any particular degree of truth, but the rumors are that it's not the process, but the architecture teams that are delaying high-end MTL. Apparently, again according to rumors, the laptop and desktop variants are worked on by different teams, and apparently the desktop team is having troubles, and so supposedly that's why Intel is launching what was intended as laptop CPUs as desktop i5s.Reply
If I had to guess, we will see meteor like for mobile, and possibly as a show off SKU for their power efficiency per core or something in the lower desktop segments. If you believe the rumor mill, the eight P core parts were dropped, probably for cost savings given the time., Which would mean they could launch lower P core parts for lower segment parts, while retaining some kind of boosted raptor lake for the new i9s. Reply
They have always been unified. The only difference is that on some generations of CPU Intel makes them mobile or desktop only, or for a period of time depending on what suits thier business needs. Production estimates, sales estimates, economics etc. Reply
Because that is the holy grail of SOC architecture, and the more flexibility you design in, the more expensive it will be to produce it. Remember that packages, interposers, VRs, and other components have limits -- physically, mechanically and electrically. Look at all the wasted space on a Zen3 with only one of the CCD populated. So right there you can see the package is way more expensive to accommodate the flexibility. Also why is there that huge gap between the IO die and the CCDs? Because the technology did not exist at the time to put them all on a passive base die for a reasonable cost. Intel waited too long to make the leap to disaggregating client SOCs, and it took them an extra 3 years to figure out what was wrong with 10nmReply
Not /quite/ Intel's first disaggregated consumer CPU: that would be Lakefield. Big/little cores, multiple dies of different process nodes, TSVs, and Foveros, all got their test run there. Reply
The CPU tile is the important tile. And Intel readily admits that TSMC has been ahead of Intel in process technology the past few years. If Intel finds it cost beneficial to use TSMC for the I/O and SOC tiles, which are produced on older nodes, what does that have to do with Meteor Lake or Intel forging ahead with their cutting edge nodes with Intel 4?
It's interesting that they prefer TSMC for the GPU tiles, though. I wonder if that will change once they are producing in volume on 18A. I say that because Intel seems to want to compete with TSMC for Nvidia's orders. Anyway, get used to it, because Intel plans to use other foundries for various parts whenever it makes sense to do so. Intel has been one of TSMC's largest customers and will likely continue to be.Reply
The reason which fab is doing what matters is - GPUs care mainly about density - CPUs (in the intel design style) care about high frequency, power be damned
The MTL split suggests that i4 can't achieve high density. Or can't achieve it at decent prices, Which is significant for both foundry (most customers skew more to prioritizing density over frequency) and for the expected cost of future Intel devices (which will be competing against both AMD on TSMC and various ARM on various foundries).
The whole point of The Innovator's Dilemma, a point which old Intel understood well, is that you cannot save your business by retreating upmarket, selling products are somewhat better performing, yes, but a LOT more expensive. Didn't work for IBM, didn't work for DEC, didn't work for SGI or SUN. (Kinda works for Apple, but they are slightly more expensive, not a LOT more, and provide an associated bundle of functionality that people consider worth the cost; they're not selling a commodity the way Intel is.)
Intel seems destined to follow the IBM track of being able to boast "#1 fastest" but at prices almost no-one is willing to pay... Certainly every new detail we learn suggests that Intel's next designs may be that few percent faster than AMD; but if that requires twice the cost (which is where all the evidence points), well, AMD, get ready for buying all the TSMC capacity Apple hasn't already snarfed up...Reply
Your first statement is right. Intel 4 doesn't support HD libraries. It's a client only node purely built for performance only at the cost of density. Not suitable for GPUs. Intel's first pure intel foundry based GPU is probably Druid based on 18A.Reply
Does it matter? Every currently known end user device type and general cloud computing server is heading for ARM based infrastructure anyways. We are past the point where the software made a switch impossible, so it’s only a matter of years before everything mainstream is ARM because of the power envelope and cost. x86 is deathspiraling for the volume market.Reply
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
18 Comments
Back to Article
TristanSDX - Thursday, April 27, 2023 - link
How this play with rumored / leaked Raptor Lake Refresh release ? Intel is silent about that refresh, and other leaks suggest MTL for desktop to be launched anyway.For Intel launch mean 'silicon out of factory' not 'in shops' Reply
Dolda2000 - Thursday, April 27, 2023 - link
Well, the rumors are that MTL will be used primarily for laptop, and *possibly* also for some i5 variants. Not saying that's what's going to happen, but that's what the rumor mill is saying. Replyikjadoon - Sunday, May 7, 2023 - link
So yields are still rather poor and this is just getting pushed out the door to meet executive timelines.— 1st 14nm CPU was basically only mobile (Broadwell)
— 1st 10nm CPU was basically only mobile (Ice Lake and Cannon Lake)
— 1st 7nm (Intel 4) CPU will be primarily mobile
See, TSMC does the same ramp (~90 to 110 mm^2 iPhone chips first), but they're humble enough to admit it as normal and part of the process.
Intel? "Oh, God, no. QUICK, a re-brand! Confuse the customers!" Reply
Dolda2000 - Friday, May 12, 2023 - link
>So yields are still rather poor and this is just getting pushed out the door to meet executive timelines.Again, not to ascribe the rumor mill any particular degree of truth, but the rumors are that it's not the process, but the architecture teams that are delaying high-end MTL. Apparently, again according to rumors, the laptop and desktop variants are worked on by different teams, and apparently the desktop team is having troubles, and so supposedly that's why Intel is launching what was intended as laptop CPUs as desktop i5s. Reply
Drumsticks - Thursday, April 27, 2023 - link
If I had to guess, we will see meteor like for mobile, and possibly as a show off SKU for their power efficiency per core or something in the lower desktop segments. If you believe the rumor mill, the eight P core parts were dropped, probably for cost savings given the time., Which would mean they could launch lower P core parts for lower segment parts, while retaining some kind of boosted raptor lake for the new i9s. Replybansheexyz - Thursday, April 27, 2023 - link
I think it's time to unify mobile and desktop chips somehow, I'm tired of waiting an extra 6 months for desktop versions. Replygoatfajitas - Friday, April 28, 2023 - link
They have always been unified. The only difference is that on some generations of CPU Intel makes them mobile or desktop only, or for a period of time depending on what suits thier business needs. Production estimates, sales estimates, economics etc. Replydrothgery - Friday, April 28, 2023 - link
It's been clear for a while that they're not doing this, but it's never been clear to me why they weren't planning on something like this:Desktop & HX i7 & i9 - 2x compute tile + small GPU tile + desktop IO tile (aka would top out at 12P + 16E)
Desktop & HX i3 & i5 - 1x compute tile + small GPU tile + desktop IO tile
H, P, and U series - 1x compute tile + big GPU tile + laptop IO tile Reply
jjjag - Monday, May 1, 2023 - link
Because that is the holy grail of SOC architecture, and the more flexibility you design in, the more expensive it will be to produce it. Remember that packages, interposers, VRs, and other components have limits -- physically, mechanically and electrically. Look at all the wasted space on a Zen3 with only one of the CCD populated. So right there you can see the package is way more expensive to accommodate the flexibility. Also why is there that huge gap between the IO die and the CCDs? Because the technology did not exist at the time to put them all on a passive base die for a reasonable cost. Intel waited too long to make the leap to disaggregating client SOCs, and it took them an extra 3 years to figure out what was wrong with 10nm Replydrothgery - Tuesday, May 2, 2023 - link
Eh, desktop and HX parts are niche anyway, and as you said it'd basically be what AMD's doing there anyway.Non-HX laptop parts (aka where the real high volume is) wouldn't have any extra space. Reply
edzieba - Friday, April 28, 2023 - link
Not /quite/ Intel's first disaggregated consumer CPU: that would be Lakefield. Big/little cores, multiple dies of different process nodes, TSVs, and Foveros, all got their test run there. Replybrucethemoose - Friday, April 28, 2023 - link
Is this the gen with a M1 Pro-esque 256-bit part, or is that Arrow Lake?I would really love an M1 Max-ish desktop part for running LLMs... without being locked to OSX or Metal. Reply
lefty2 - Friday, April 28, 2023 - link
Meteor lake and Intel 4... (yeah, but don't tell anyone that it's mostly TSMC fabbed, shuu!) ReplyYojimbo - Friday, April 28, 2023 - link
The CPU tile is the important tile. And Intel readily admits that TSMC has been ahead of Intel in process technology the past few years. If Intel finds it cost beneficial to use TSMC for the I/O and SOC tiles, which are produced on older nodes, what does that have to do with Meteor Lake or Intel forging ahead with their cutting edge nodes with Intel 4?It's interesting that they prefer TSMC for the GPU tiles, though. I wonder if that will change once they are producing in volume on 18A. I say that because Intel seems to want to compete with TSMC for Nvidia's orders. Anyway, get used to it, because Intel plans to use other foundries for various parts whenever it makes sense to do so. Intel has been one of TSMC's largest customers and will likely continue to be. Reply
name99 - Monday, May 1, 2023 - link
The reason which fab is doing what matters is- GPUs care mainly about density
- CPUs (in the intel design style) care about high frequency, power be damned
The MTL split suggests that i4 can't achieve high density. Or can't achieve it at decent prices,
Which is significant for both foundry (most customers skew more to prioritizing density over frequency) and for the expected cost of future Intel devices (which will be competing against both AMD on TSMC and various ARM on various foundries).
The whole point of The Innovator's Dilemma, a point which old Intel understood well, is that you cannot save your business by retreating upmarket, selling products are somewhat better performing, yes, but a LOT more expensive. Didn't work for IBM, didn't work for DEC, didn't work for SGI or SUN. (Kinda works for Apple, but they are slightly more expensive, not a LOT more, and provide an associated bundle of functionality that people consider worth the cost; they're not selling a commodity the way Intel is.)
Intel seems destined to follow the IBM track of being able to boast "#1 fastest" but at prices almost no-one is willing to pay... Certainly every new detail we learn suggests that Intel's next designs may be that few percent faster than AMD; but if that requires twice the cost (which is where all the evidence points), well, AMD, get ready for buying all the TSMC capacity Apple hasn't already snarfed up... Reply
SiliconFly - Saturday, May 6, 2023 - link
Your first statement is right. Intel 4 doesn't support HD libraries. It's a client only node purely built for performance only at the cost of density. Not suitable for GPUs. Intel's first pure intel foundry based GPU is probably Druid based on 18A. ReplyDer Keyser - Saturday, May 6, 2023 - link
Does it matter? Every currently known end user device type and general cloud computing server is heading for ARM based infrastructure anyways. We are past the point where the software made a switch impossible, so it’s only a matter of years before everything mainstream is ARM because of the power envelope and cost. x86 is deathspiraling for the volume market. ReplySiliconFly - Saturday, May 6, 2023 - link
AMD uses TSMC. It's not like they have a monopoly or exclusivity with TSMC. If others want to use it, they very well can. Reply